报告题目：Towards Automated Structural Inspection with Deep Learning and Unmanned Aerial Vehicles

主讲人：Vedhus Hoskere（Ph.D candidate, Civil Engineering, University of Illinois-Urbana Champaign, USA）

时  间：2018年8月14日14:00—15:00

地  点：土木工程学院402室

Abstract:

Current methods of practice for inspection of civil infrastructure typically involve visual assessments conducted manually by trained inspectors. Manual inspections have proven to be time consuming, laborious, and prone to inaccuracies due to their subjective nature. In this seminar, we define condition-aware models as a 3D mesh represetnataion of a structure where damage and context labels are known for each mesh cell. Condition-aware models can be a useful tool on which autoamted higher-level structurtal assessments can be built. This seminar will desribe a framework to develop condition-aware models that a enable automated inspections. The framework involves three steps (i) identification of damage (ii) idnetification of context (iii) projection of damage and context labels on to a 3D mesh model. Deep semantic segmentaiton using fully convolutional neural networks (FCNs) is proposed to conduct the first two steps. For the first step of damage detection, a general database incorporating six damage types including cracks, spalling, exposed rebar, corrosion, fatigue cracks and asphalt cracks is created for training the proposed deep networks. Netowork configurations are proposed and tested for pixel-wise accuracy on unseen images. Then, to deomnstrte the development of condition-aware models, the specific application of post-earthquake building inspections is considered. For building structures, another database is created for contextual information such as the presence of a building and visually identifiable components like windows and doors. A similar semantic segmentaiton approach is employed to identify context labels. Then, the structure-from motion (SFM) algorithm is employed with multi-view data from a single structure to create a 3D mesh. The trained deep models are used to infer the damage and context on the multi-vew dataset. The inferred labels are projected onto the 3D mesh to generate the condition-aware model. The proposed framework is implemented on video acquired after the 2017 central Mexico earthquake in downtown Mexico City using unmanned aerial vehicles. The results demonstrate the promise of the proposed approach as a critical step in enabling automated structural inspections.